Welcome back and give it a try!

Other people were saying they were having quality issues when training db characters on illustrious. No harm in trying though. You always had the best db loras so you can probably pull it off.

@Jval33 Me, it was me, look for another Pan, is bad. And my Tenshinhan is horrible, I don't even want to upload.

Now, your higher dimension might let the rest of the style bleed a bit more and might help, so let me know please.

@IndolentCat What kind of dimension settings have you tried? I've only made one lora so far, and that was for SD 1.5 (which I still use for inpainting sometimes), and I used 16 on that one.

sure and welcome back King!

@LemonSparkle The issue is Dragon Ball characters on Illustrious, and that's with screencaps/renders, any character that you can do with booru images, I'm sure it might work. Not training in general. My go to settings on SDXL are 8/4. On 1.5 it was 32/16. That's for characters.

@IndolentCat Really it works on just 4 or 8 for SDXL? So that's not too low? I was wondering if I should make an improved version of my lora, maybe a version for a higher res base model, (because I made 1024 versions of my whole dataset first, then scaled to 512 for it), but I'm still not really sure on all the settings, I just read something that said 16 worked and above that for characters was pretty much overkill and made overly large files, but I'm not well versed on the technical side.

@LemonSparkle When I say 8/4 that means, dimension 8, alpha 4. Dimension 4 might be too low. But yeah, it works like that on SDXL.

@IndolentCat Oh okay, yeah for my LoRA I used Dimensions 16 and Alpha 8, trained on 62 images at 512x512 res, with lots of tags (idk what of that besides Dim makes a difference or not), I ended up at about 18mb in file size which didn't seem too bad, I think it works better than the other SD 1.5 version for the same character someone else made, I was pretty careful about my training images though and did a lot of edits and redraws to improve them before using, I've tried to do my best reading about it but it's like one comment I saw here that described the articles on making LoRAs here mostly come in two flavors, either its "this is what I do, but I don't know why it works", or otherwise it's like trying to translate an ancient manuscript filled with mysterious incomprehensible runes.

@LemonSparkle Yeah, is difficult to find everyday language info, the resolution above 512 might give you more details/quality, but it can also conflict with the fact that 1.5 wasn't intended for more, in theory you'd have to generate bigger too.

@TheBlackPrince My best bet at this time is to train an accompanying style (like DB/Z style), but that defeats the purpose of training just a character IMO; or diversifying the datasets, although some characters don't have enough out of style fanart for that.

@TheBlackPrince That's still miles better than my Tenshinhan 😂. Is so bad that makes me think than just drawing a couple of months I could do better. I'll try higher dim.